Train a LoRA,
then generate with it on the same API.

Upload your own images and get back a trained style LoRA registered to your account. Generate with it through the same Runware API you already use.

Pick your model
Up to 30% lessto train your models
Own the weightsYours to download
Style consistency

See the style hold.

Train once, and the same style shows up across everything you generate.

Trained on this.
  • Training image 1
  • Training image 2
  • Training image 3
  • Training image 4
  • Training image 5
  • Training image 6
Generated after training.
Generated example 1
Generated example 2
Generated example 3
Generated example 4
How it works

One call to train, one to generate.

Zip at least ten images. Optionally add a caption file next to any image and pick a trigger word to invoke the style later. Submit a training task through the endpoint and SDKs you already use. A webhook or polling request tells you when it is done.

Training task
{
  "taskType": "training",
  "model": "runware:flux-1-dev@style-lora-training",
  "inputs": { "dataset": "https://your-bucket/style-dataset.zip" },
  "importModel": {
    "air": "yourorg:house-style@1",
    "name": "House Style v1"
  },
  "webhookURL": "https://yourapp.com/hooks/training"
}

When training finishes, the LoRA is already a private model in your account. Reference it in a normal image request.

Image request
{
  "taskType": "imageInference",
  "model": "runware:101@1",
  "lora": [{ "model": "yourorg:house-style@1" }],
  "positivePrompt": "product shot of a ceramic mug, house-style"
}
Iterate without starting over

Pick up where you left off.

Train for a few hundred steps, check the result, then continue training the same checkpoint instead of starting from zero. Dial the style in over several passes, and only pay for the extra steps you run.

Training result after pass 1
Pass 1
Pass 2
Pass 3
Continue from checkpoint
Compute used400 steps
Ownership

Your model is yours.

Private by default.

Nobody else can see or use your trained model.

Download the weights.

Standard .safetensors, ready to load in any tool that supports LoRAs, or keep it hosted with us.

We never train on your data.

Uploaded datasets and trained models are never used to improve our own models.

Datasets deleted after training.

Your images are removed once the job completes.

Built to build on

Put training inside your own product.

If your product lets your users create their own styles, an avatar app or a creative community for example, this is the training backend so you do not have to run one. Your users see your wizard. We handle the queue and the GPUs.

Async API
Webhooks
Machine-readable errors
Per-task logs
Pricing

Pay for the training you use.

You pay for the compute you use, and nothing else. Here is where cost sits, and where it does not.

Upload

Free

Validate

Free

Train

Pay for the steps you run

Own

Free. The model is yours

Run

Pay per generation

If validation fails, training never starts and you pay nothing.

Cheaper to train

up to 30% less

than the industry to train a LoRA

Cheaper to run

up to 90% less

than the industry to generate

Priced per step

You pay per training step, and you set the number of steps. That means you can work out the cost before the job runs, not after it finishes.

Savings vary by base model, up to 90%.

FAQ

Questions, answered.

A small add-on trained on your own images that teaches an existing model a specific style, without retraining the whole model. It trains on a small dataset and applies at generation time.

LoRA training

Start training on Runware.

Train a style LoRA and generate with it on the same API. Talk to us if you are planning high volumes.

View modelsTalk to us